Token Economics: Why Pricing Is Per Token

Every major LLM API charges per token, and per-token cost differs between input and output. This is not arbitrary — it directly tracks the compute the provider must spend.

Model (mid-2026 pricing)	Input ($/1M)	Output ($/1M)
GPT-4o	$2.50	$10.00
Gemini 2.5 Pro	$1.25	$10.00
Claude 3.7 Sonnet	$3.00	$15.00
Gemini 2.5 Flash	$0.30	$2.50

Why output is 4-10× more expensive than input

Input tokens are processed in parallel during the prefill phase — one big matmul over the entire prompt. Output tokens are produced sequentially, one full forward pass per token through every layer, attending to the entire growing context. Per-token compute is similar; the per-token throughput on shared hardware differs by an order of magnitude. That gap is what the pricing reflects.

Implications for your application: long context is cheaper than long generation. RAG-style "stuff lots of retrieved documents in, generate a short answer" is favorably priced; "stuff a short prompt in, generate a 5,000-token essay" is the expensive shape. When you architect features, understanding this asymmetry shifts which strategies are economic.

Code

Cost estimator — be honest before you build·python

def estimate_cost(prompt_tokens, output_tokens,
                  in_per_1m, out_per_1m):
    return prompt_tokens / 1_000_000 * in_per_1m \
         + output_tokens / 1_000_000 * out_per_1m

# Pricing (USD / 1M tokens) — update with current sheet
PRICES = {
    "gpt-4o":             (2.50, 10.00),
    "claude-3.7-sonnet":  (3.00, 15.00),
    "gemini-2.5-flash":   (0.30,  2.50),
}

# Daily budget for 100k requests, average 800-token prompt + 400-token answer
for model, (in_p, out_p) in PRICES.items():
    daily = 100_000 * estimate_cost(800, 400, in_p, out_p)
    print(f"{model:>22}  ${daily:>9,.2f}/day")

Token Economics: Why Pricing Is Per Token

Why output is 4-10× more expensive than input

Code

External links

Exercise

Progress

Comments 0